Remote sensing ground object knowledge graph construction method based on cooperation of geographic object geometric features and graph enhancement retrieval

Abstract

The invention discloses a remote sensing ground object knowledge graph construction method based on geographic object geometric features and graph enhancement retrieval collaboration. The method comprises the following steps: preparing a data set; basic triple extraction and feature analysis are carried out; semantic and relation enhancement assisted by LLM; flexible quantification and output of knowledge of the ground object knowledge graph; and correcting knowledge conflicts and constructing a final edition map. According to the method, a tensor decomposition model is used for training a ground feature knowledge graph to obtain a knowledge graph embedded vector, through an unsupervised / semi-supervised mode, low-dimensional vector representation of ground feature categories is learned from a knowledge graph triple, the knowledge graph embedded vector is used for embedding a semantic relationship between categories, global semantic priori knowledge is provided for a downstream model, and the global semantic priori knowledge is provided for the downstream model. And the semantic confusion problem in the remote sensing image is effectively solved.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph technology, and in particular to a method for constructing remote sensing ground feature knowledge graphs based on the synergy of geographic object geometric features and graph-enhanced retrieval. Background Technology

[0002] Knowledge graphs (KGs) have become a core technology for managing and utilizing large-scale, complex data. However, in the fields of remote sensing and geographic information, which have significant geospatial distribution characteristics, existing technologies face the following prominent problems:

[0003] 1. High reliance on manual annotation and high cost: Traditional knowledge graph construction methods, especially those based on supervised learning, require large-scale, high-quality labeled datasets. In the field of geographic information, acquiring such data is not only highly challenging and extremely costly, but also requires the deep involvement of scarce domain experts, making annotation a bottleneck in knowledge graph construction.

[0004] 2. Weak generalization ability in low-resource scenarios: In scenarios with sparse labeled data and few or zero samples, existing models have difficulty effectively identifying and extracting new entities or relationships that have not been seen before, resulting in insufficient robustness.

[0005] 3. Rigid Spatial Relationships: Traditional topological relationships (such as "adjacent" and "containment") are binary rigid logic (yes / no), which is difficult to express the fuzzy and gradual geographical relationships in the real world. This rigidity makes it difficult for downstream remote sensing analysis tasks to use accurate spatial prior knowledge for correction, easily leads to discontinuous classification boundaries, misjudgments of complex scenes, and severely limits the topological consistency and spatial quality of segmentation results. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a method for constructing a remote sensing geographic feature knowledge graph based on the synergy of geographic object geometric features and graph augmentation retrieval, which is simple in algorithm, low in cost, and robust.

[0007] The technical solution of this invention to solve the above-mentioned technical problems is: a method for constructing a remote sensing ground feature knowledge graph based on the synergy of geographic object geometric features and graph-enhanced retrieval, comprising the following steps:

[0008] Step S1, Dataset Preparation: Prepare remote sensing image data, metadata information for the project dataset, and ground truth label dataset for remote sensing images. External corpora and primary knowledge sources;

[0009] Step S2, Basic Triple Extraction and Feature Parsing: Using the metadata of the project dataset and the coarse segmentation results of remote sensing images, identify geographic objects (Geo-Objects) and their geometric and spectral features, and construct a basic geo-KG (Geographic Knowledge Graph).

[0010] Step S3, LLM-assisted semantic and relational enhancement: Introducing a large-scale language model (LLM), based on an external corpus, and utilizing... The mechanism retrieves structured evidence chains through topic community recall and multi-hop link retrieval, and performs knowledge expansion, relationship completion, and semantic verification on the basic geo-KG knowledge graph.

[0011] Step S4, Flexible Quantization and Output of Geographic Knowledge Graph Knowledge: Using Tensor Decomposition Model Training the Geo-KG knowledge graph yields the knowledge graph embedding vectors. Based on statistics from a large-scale remote sensing ground truth dataset, a probabilistic topological intensity matrix was obtained through quantification. ;

[0012] Step S5, Knowledge Conflict Correction and Final Graph Construction: Combining primary knowledge sources, through... Evidence chains are retrieved for conflicting knowledge, and LLM combines knowledge aggregation and conflict resolution strategies to adjudicate conflicts, ensuring that the final output is a complete knowledge graph of remote sensing features in the vertical domain.

[0013] The above-mentioned method for constructing a remote sensing knowledge graph based on the synergy of geographic object geometric features and graph-enhanced retrieval, specifically step S2, involves the following process:

[0014] Step S21, coarse segmentation: Apply the SLIC algorithm to the remote sensing image to generate Geo-Object candidate regions with preliminary spatial boundaries;

[0015] Step S22, Metadata Fusion: Associate the Geo-Object candidate region with the metadata information of the project dataset to ensure that each superpixel block obtains a preliminary semantic label and spatial context, and complete the preliminary semantic recognition of the Geo-Object.

[0016] Step S23, Geometric feature calculation: Calculate the geometric features of each obtained Geo-Object to quantify the shape and size attributes of the land cover;

[0017] Step S24, Spectral feature calculation: Statistical analysis is performed on all pixels within the Geo-Object candidate region to calculate the spectral features of the Geo-Object.

[0018] Step S25: Construct the basic triplet.

[0019] The aforementioned method for constructing a remote sensing knowledge graph based on the synergy of geographic object geometric features and graph-enhanced retrieval, specifically step S25, involves: using the preliminary classification labels of geographic objects (Geo-Objects) as entities to determine the node types of the knowledge graph; and then using a dimensional extension nine-intersection model... Spatial operations are performed to extract initial topological relationships; initial functional relationships are extracted based on weakly supervised semantic initial rules; all calculated geometric and spectral features are used as attributes of Geo-Object entities, and finally, the initial set of basic triples for the geographic feature knowledge graph is output. , ,in Representative attribute triples, This represents the union operation. Indicates the head entity. Indicates the tail entity. Indicates a relationship.

[0020] The above-mentioned method for constructing a remote sensing knowledge graph based on the synergy of geographic object geometric features and graph-enhanced retrieval, specifically step S3, involves the following process:

[0021] Step S31, Uncertainty Analysis and Identification: For The confidence level of the triples is evaluated, and triples with confidence levels below a preset threshold or lacking semantic evidence are selected for enhancement.

[0022] Step S32: Set up the GraphRAG deep retrieval mechanism: Based on the head entity, tail entity, and relation type of the triple to be enhanced, construct a refined retrieval query; utilize a semantic graph pre-built from an external corpus; identify semantic clusters using the Leiden algorithm; recall the entire semantic cluster related to the target topic; perform multi-hop path analysis to mine implicit causal and semantic chains across paragraphs and documents; rank the recalled text blocks by relevance; select the most informative text context set to form a chain of evidence. ;

[0023] Step S33, Structured Knowledge Extraction and Completion Based on LLM: LLM for Evidence Chain To conduct analysis, The semantic rationality of the relationships in the dataset is verified. The LLM process performs entity extraction, relation extraction, attribute extraction, and event extraction tasks, producing formalized knowledge triples and event quadruples. The LLM identifies synonymous entities, assigns them unified labels, and generates new functional relation triples that conform to the common sense of the vertical domain. Finally, the enhanced Geo-KG set is output. .

[0024] The above-mentioned method for constructing a remote sensing knowledge graph based on the synergy of geographic object geometric features and graph-enhanced retrieval, in step S4, utilizes a tensor decomposition model. Training the Geo-KG knowledge graph yields the knowledge graph embedding vectors. The process is as follows:

[0025] First, the enhanced Geo-KG collection All semantic triples in the training set , ,in It is a head entity. , It is a collection of entities. It's a relationship. , It is a set of semantic relations. It is a tail entity. ;

[0026] Then, model selection and initialization were performed, using a tensor decomposition model. As a knowledge graph embedding algorithm, initialize all entities and relationships Complex vector embeddings;

[0027] Then minimize the scoring function During training, the scoring function is defined as follows:

[0028]

[0029] in, yes The conjugate of the set is used, and the training objective is to maximize the score of correct triples and minimize the score of incorrect triples. The real part extraction operator is , and <> is the trilinear dot product over the complex field;

[0030] After training, the knowledge graph embedding vectors of all categories of Geo-Object are obtained. , , For discrete entity indexing, For the first The latent feature vectors of each entity represent global semantic association information. For entity collection The base number.

[0031] In the aforementioned method for constructing a remote sensing knowledge graph based on the synergy of geographic object geometric features and graph-enhanced retrieval, step S4 involves quantizing the probabilistic topological intensity matrix based on statistics from a large-scale remote sensing ground truth dataset. The process is as follows:

[0032] First, use the ground truth label dataset of remote sensing images. The set of land cover categories is , , For the first Each category of land cover It represents the total number of land cover categories;

[0033] Then, spatial co-occurrence frequency statistics are performed, which includes pixel pair statistics, total pixel pair statistics, and calculation of the initial co-occurrence frequency. Pixel pair statistics are... Each pair of land cover categories , For the first Each category of land cover For the first Each land cover category, statistics and The number of spatially adjacent or functionally symbiotic pixel pairs Total pixel pair statistics are statistics Number of all pixel pairs appearing ; Calculate the first co-occurrence frequency in the initial co-occurrence frequency matrix Line number Column elements :

[0034]

[0035] Next, the probabilistic topological strength value is calculated, and LLM and Geo-Object geometric factors are introduced for correction, resulting in the final strength value. The mixture is weighted and fused from three factors and then processed by the Sigmoid function. Normalization yields:

[0036]

[0037] in, It is a weighted fusion score, defined as:

[0038]

[0039] in, For geometric feature factors based on Geo-Object, Based on Statistical factors of the knowledge graph structure. This serves as a verification factor for the semantic rationality of topological relationships based on LLM. All are weighted coefficients;

[0040] Finally, output the probabilistic matrix, obtaining a dimension of Probabilized topological strength matrix , As a soft constraint for topological consistency in downstream models.

[0041] The above-mentioned method for constructing a remote sensing knowledge graph based on the synergy of geographic object geometric features and graph-enhanced retrieval, in step S5, involves... The process of evidence chain retrieval for conflicting knowledge is as follows:

[0042] First, , And an inviolable first-level knowledge source as input, the first-level knowledge source being a hard rule base;

[0043] Then, semantic embedding verification, topology strength verification, and domain logic verification are performed respectively to identify conflicts;

[0044] Finally, all identified conflicts are marked as a set of triples to be corrected. ,use right For each conflicting triple, a chain of evidence is retrieved. The goal is to obtain a structured, traceable chain of evidence. Chain of evidence Points to the most authoritative text blocks or logical reasoning chains in an external corpus, used to support or refute conflicting knowledge.

[0045] The aforementioned method for constructing a remote sensing geographic feature knowledge graph based on the synergy of geographic object geometric features and graph-enhanced retrieval, in step S5, aims to ensure that the semantic relationships between geographic objects conform to domain common sense. The semantic embedding verification process involves: evaluating... and Vector distance between ,like Less than the preset threshold However, the definition of a first-level knowledge source and If they must be independent, then they are marked as semantic conflicts;

[0046] The goal of topology strength verification is to ensure that spatial symbiotic flexibility conforms to hard rules. The process of topology strength verification is as follows: Check The intensity value, if defined by the first-level knowledge source and They should never coexist, but the verification process found... Greater than the set threshold If so, it is marked as a topological conflict;

[0047] The goal of domain logic verification is to ensure that the graph structure conforms to the domain axioms. The process of domain logic verification is as follows: Using triples in the equation for rule-based logical reasoning, if... There exists The definition of a first-level knowledge source It is impossible to be If it is included, it is marked as a logical conflict.

[0048] The above-mentioned method for constructing a remote sensing knowledge graph based on the synergy of geographic object geometric features and graph-enhanced retrieval, in step S5, the process of LLM combining knowledge aggregation and conflict resolution strategies to adjudicate conflicts is as follows:

[0049] First, LLM receives First-level knowledge source rules and chain of evidence ;

[0050] Then, the LLM executes the following knowledge aggregation and conflict resolution strategies:

[0051] Credibility assessment: Based on the authority of the evidence source Conduct credibility scoring;

[0052] Duplicate entity disambiguation: Re-verify whether there are synonymous unmerged entities in the conflicting triples;

[0053] Relation consistency check: Check whether the relation logic of conflicting triples conforms to domain axioms and authoritative evidence;

[0054] Finally, LLM correction Incorrect semantic embedding values ​​or topological strength values ​​are used to generate the final map set. .

[0055] The above-mentioned method for constructing a remote sensing knowledge graph based on the synergy of geographic object geometric features and graph-enhanced retrieval, in step S5, ultimately, will... Write the data into the Neo4j graph database and output a vertical domain remote sensing feature knowledge graph that is compatible with the corresponding field or project, including... , , ,Will As external knowledge, reconstruct semantic clusters to improve the next round. This improves the accuracy of recall and forms a closed-loop knowledge mining system.

[0056] The beneficial effects of this invention are as follows:

[0057] 1. This invention utilizes a tensor decomposition model. Training the Geo-KG knowledge graph yields the knowledge graph embedding vectors. Through unsupervised / semi-supervised methods, utilizing Learn low-dimensional vector representations of land cover categories from knowledge graph triples. It embeds semantic relationships between categories, providing global semantic prior knowledge for downstream models and effectively solving the semantic confusion problem in remote sensing images.

[0058] 2. This invention is based on the statistical analysis of a large-scale remote sensing ground truth dataset, and quantifies to obtain a probabilistic topological intensity matrix. Transforming traditional binary topological relationships into Probability intensity values ​​within the interval, probabilistic topological intensity matrix The intensity of flexible symbiosis between land cover categories in spatial coexistence or repulsion was quantified, serving as a soft constraint for topological consistency in downstream deep learning models. Attached Figure Description

[0059] Figure 1 This is the overall flowchart of the present invention.

[0060] Figure 2 This is a flowchart of the basic triplet extraction and feature parsing process in this invention.

[0061] Figure 3 This is a flowchart of the LLM-assisted semantic and relational enhancement in this invention.

[0062] Figure 4 This is a diagram illustrating the calculation model of the probabilistic topological strength matrix in this invention. Detailed Implementation

[0063] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0064] like Figure 1 As shown, a method for constructing a remote sensing geographic feature knowledge graph based on the synergy of geographic object geometric features and graph-enhanced retrieval includes the following steps:

[0065] Step S1, Dataset Preparation: Prepare remote sensing image data, metadata information for the project dataset, and ground truth label dataset for remote sensing images. External corpora and primary knowledge sources.

[0066] Step S2, Basic Triple Extraction and Feature Parsing: Using the metadata of the project dataset and the coarse segmentation results of remote sensing images, identify geographic objects (Geo-Objects) and their geometric and spectral features, and construct a basic geo-KG knowledge graph.

[0067] like Figure 2 As shown, the specific process of step S2 is as follows:

[0068] Step S21, coarse segmentation: Apply the SLIC algorithm to the remote sensing image to generate Geo-Object candidate regions with preliminary spatial boundaries; the SLIC algorithm effectively generates compact and uniform superpixel blocks by clustering in five-dimensional space, which serve as the basis for subsequent Geo-Objects.

[0069] Step S22, Metadata Fusion: Associate the Geo-Object candidate region with the metadata information of the project dataset (including but not limited to geographic coordinates and preliminary classification labels) to ensure that each superpixel block obtains a preliminary semantic label and spatial context, and completes the preliminary semantic recognition of the Geo-Object.

[0070] Step S23, Geometric feature calculation: Calculate the geometric features of each obtained Geo-Object. Geometric features include, but are not limited to, area, perimeter, shape index (length-to-width ratio, etc.), and compactness (perimeter squared / area ratio), which are used to quantify the shape and size attributes of the land feature.

[0071] Step S24, Spectral feature calculation: Statistical analysis is performed on all pixels within the Geo-Object candidate region to calculate the spectral features of the Geo-Object. The spectral features include the mean, standard deviation (measures the spectral uniformity within the region), and specific band combination indices (Normalized Difference Vegetation Index NDVI, texture, and others).

[0072] Step S25: Construct the basic triplet.

[0073] The initial classification labels of geo-objects (e.g., buildings, roads) are used as entities to determine the node types of the geo-feature knowledge graph; a nine-intersection model based on dimensional expansion is then used. Spatial operations are performed to extract initial topological relationships; initial functional relationships are extracted based on weakly supervised semantic initial rules; all calculated geometric and spectral features are used as attributes of Geo-Object entities, and finally, the initial set of basic triples for the geographic feature knowledge graph is output. , ,in Representative attribute triples, This represents the union operation. Indicates the head entity. Indicates the tail entity. Indicates a relationship.

[0074] Step S3, LLM-assisted semantic and relational enhancement: Introducing a large-scale language model (LLM), based on an external corpus, and utilizing... The mechanism retrieves structured evidence chains through topic community recall and multi-hop link retrieval, and performs knowledge expansion, relationship completion, and semantic verification on the basic geo-KG knowledge graph.

[0075] like Figure 3 As shown, the specific process of step S3 is as follows:

[0076] Step S31, Uncertainty Analysis and Identification: For The confidence level of the triples is evaluated, and triples with confidence levels below a preset threshold or lacking semantic evidence are selected for enhancement.

[0077] Step S32: Set up the GraphRAG deep retrieval mechanism: Based on the head entity, tail entity, and relation type of the triple to be enhanced, construct a refined retrieval query; utilize a semantic graph pre-built from an external corpus; identify semantic clusters using the Leiden algorithm; recall the entire semantic cluster related to the target topic; perform multi-hop path analysis to mine implicit causal and semantic chains across paragraphs and documents; rank the recalled text blocks by relevance; select the most informative text context set to form a chain of evidence. This ensures that the external knowledge input into the LLM is of high quality and has broad coverage;

[0078] Step S33, Structured Knowledge Extraction and Completion Based on LLM: LLM for Evidence Chain To conduct analysis, The semantic rationality of the relationships in the dataset is verified. The LLM process performs entity extraction, relation extraction, attribute extraction, and event extraction tasks, producing formalized knowledge triples and event quadruples. The LLM identifies synonymous entities, assigns them unified labels, and generates new functional relation triples that conform to the common sense of the vertical domain. Finally, the enhanced Geo-KG set is output. .

[0079] Step S4, Flexible Quantization and Output of Geographic Knowledge Graph Knowledge: Using Tensor Decomposition Model Training the Geo-KG knowledge graph yields the knowledge graph embedding vectors. Based on statistics from a large-scale remote sensing ground truth dataset, a probabilistic topological intensity matrix was obtained through quantification. .

[0080] Using tensor decomposition model Training the Geo-KG knowledge graph yields the knowledge graph embedding vectors. The process is as follows:

[0081] First, the enhanced Geo-KG collection All semantic triples in the training set , ,in It is a head entity. , It is a collection of entities. It's a relationship. , It is a set of semantic relations. It is a tail entity. ;

[0082] Then, model selection and initialization were performed, using a tensor decomposition model. As a knowledge graph embedding algorithm, initialize all entities and relationships Complex vector embeddings;

[0083] Then minimize the scoring function During training, the scoring function is defined as follows:

[0084]

[0085] in, yes The conjugate of the set is used, and the training objective is to maximize the score of correct triples and minimize the score of incorrect triples. The real part extraction operator transforms complex vector interactions into a scalar score, used to quantify the authenticity or confidence of triples in the latent semantic space; 〈〉 represents the complex-domain trilinear dot product, which, through the introduction of conjugate transformation, enables the model to effectively capture asymmetric features in geospatial relationships (i.e., distinguish between...). and The directional difference is the core operator for achieving high-precision semantic embedding;

[0086] After training, the knowledge graph embedding vectors of all categories of Geo-Object are obtained. , , For discrete entity indexes, it involves traversal. In the algorithm implementation, the discrete index variables... As a unique identifier, it is used to locate a specific first in the map. A geographic object or semantic concept node; For the first Potential feature vectors of each entity; For entity collection The cardinality represents the total number of unique nodes contained in the constructed geographic feature knowledge graph. The total number of unique nodes defines the... The scale, that is, the total capacity of knowledge.

[0087] It incorporates global semantic prior knowledge between categories.

[0088] Based on statistics from a large-scale remote sensing ground truth dataset, a probabilistic topological intensity matrix is ​​obtained through quantization. The process is as follows:

[0089] First, a large-scale, high-reliability remote sensing image ground truth label dataset is used. The set of land cover categories is , , For the first Each category of land cover It represents the total number of land cover categories;

[0090] Then, spatial co-occurrence frequency statistics are performed, which includes pixel pair statistics, total pixel pair statistics, and calculation of the initial co-occurrence frequency. Pixel pair statistics are... Each pair of land cover categories , For the first Each category of land cover For the first Each land cover category, statistics and The number of spatially adjacent or functionally symbiotic pixel pairs Total pixel pair statistics are statistics Number of all pixel pairs appearing ; Calculate the first co-occurrence frequency in the initial co-occurrence frequency matrix Line number Column elements :

[0091]

[0092] Next, the probabilistic topological strength value is calculated, such as Figure 4 As shown, LLM and Geo-Object geometric factors are introduced for correction, resulting in the final intensity value. The mixture is weighted and fused from three factors and then processed by the Sigmoid function. Normalization yields:

[0093]

[0094] in, It is a weighted fusion score, defined as:

[0095]

[0096] in, For geometric feature factors based on Geo-Object, Based on Statistical factors of the knowledge graph structure. This serves as a verification factor for the semantic rationality of topological relationships based on LLM. All are weighted coefficients;

[0097] Finally, output the probabilistic matrix, obtaining a dimension of Probabilized topological strength matrix , As a soft constraint for topological consistency in downstream models.

[0098] Step S5, Knowledge Conflict Correction and Final Graph Construction: Combining primary knowledge sources, through... Evidence chains are retrieved for conflicting knowledge, and LLM combines knowledge aggregation and conflict resolution strategies to adjudicate conflicts, ensuring that the final output is a complete knowledge graph of remote sensing features in the vertical domain.

[0099] pass The process of evidence chain retrieval for conflicting knowledge is as follows:

[0100] First, , And an inviolable first-level knowledge source as input, the first-level knowledge source being a hard rule base;

[0101] Then, semantic embedding verification, topology strength verification, and domain logic verification are performed respectively to identify conflicts;

[0102] The goal of semantic embedding validation is to ensure that the semantic relationships between geo-objects conform to domain common sense. The process of semantic embedding validation is as follows: evaluation and Vector distance between ,like Less than the preset threshold However, the definition of a first-level knowledge source and If they must be independent, then they are marked as semantic conflicts;

[0103] The goal of topology strength verification is to ensure that spatial symbiotic flexibility conforms to hard rules. The process of topology strength verification is as follows: Check The intensity value, if defined by the first-level knowledge source and Never coexist (e.g., (Should be 0), but the verification program found... Greater than the set threshold If so, it is marked as a topological conflict;

[0104] The goal of domain logic verification is to ensure that the graph structure conforms to the domain axioms. The process of domain logic verification is as follows: Using triples in the equation for rule-based logical reasoning, if... There exists The definition of a first-level knowledge source It is impossible to be If it is included, it is marked as a logical conflict.

[0105] Finally, all identified conflicts are marked as a set of triples to be corrected. ,use right For each conflicting triple, a chain of evidence is retrieved. The goal is to obtain a structured, traceable chain of evidence. Chain of evidence Points to the most authoritative text blocks or logical reasoning chains in an external corpus, used to support or refute conflicting knowledge.

[0106] The process by which LLM combines knowledge aggregation and conflict resolution strategies to adjudicate conflicts is as follows:

[0107] First, LLM receives First-level knowledge source rules and chain of evidence ;

[0108] Then, the LLM executes the following knowledge aggregation and conflict resolution strategies:

[0109] Credibility assessment: Based on the authority of the evidence source Conduct credibility scoring;

[0110] Duplicate entity disambiguation: Re-verify whether there are synonymous unmerged entities in the conflicting triples;

[0111] Relation consistency check: Check whether the relation logic of conflicting triples conforms to domain axioms and authoritative evidence;

[0112] Finally, LLM correction Incorrect semantic embedding values ​​or topological strength values ​​are used to generate the final map set. .

[0113] Ultimately, Write the data into the Neo4j graph database and output a vertical domain remote sensing feature knowledge graph that is compatible with the corresponding field or project, including... , , ,Will As external knowledge, reconstruct semantic clusters to improve the next round. This improves the accuracy of recall and forms a closed-loop knowledge mining system.

Claims

1.A method for constructing a remote sensing geographic object knowledge graph based on cooperation of geographic object geometric features and map enhanced retrieval, characterized in that, The method comprises the following steps: Step S1, dataset preparation: prepare remote sensing image data, metadata information of project dataset, remote sensing image ground truth label dataset external corpus and primary knowledge source; Step S2, basic triple extraction and feature analysis: identifying geographical objects Geo-Object and their geometric features and spectral features by using the metadata of the project dataset and the coarse segmentation result of the remote sensing image, and constructing a basic Geo-KG; Step S3, LLM-assisted semantic and relationship enhancement: Introduce a large language model (LLM) based on external corpus, use mechanism, through topic community recall and multi-hop link retrieval, retrieve structured evidence chains, and expand the knowledge, complete the relationship, and verify the semantics of the basic Geo-KG. Step S4, flexible quantization and output of the feature knowledge graph knowledge: using tensor decomposition model Training the feature knowledge graph Geo-KG to obtain a knowledge graph embedding vector And based on the large-scale remote sensing true value data set statistics, the probability topological intensity matrix is obtained by quantization ; Step S5, knowledge conflict correction and final map construction: combined with the first level knowledge source, through The conflict knowledge is searched by evidence chain, and the LLM combines the knowledge summary and conflict resolution strategy to make a decision on the conflict, so as to ensure that the final output is a complete vertical field remote sensing feature knowledge graph. 2.The method of claim 1, wherein, The specific process of step S2 is: Step S21, coarse segmentation: applying the SLIC algorithm to the remote sensing image to generate Geo-Object candidate regions with preliminary spatial boundaries; Step S22, metadata fusion: associating the Geo-Object candidate regions with the metadata information of the project dataset to ensure that each superpixel block obtains a preliminary semantic label and spatial context, and completing the preliminary semantic recognition of the geographical objects Geo-Object; Step S23, geometric feature calculation: calculating the geometric features of each geographical object Geo-Object obtained for quantifying the shape and size properties of the features; Step S24, spectral feature calculation: performing statistics on all pixels in the Geo-Object candidate region to calculate the spectral features of the geographical objects Geo-Object; Step S25, constructing basic triples. 3.The method of claim 2, wherein, The step S25 specifically includes: taking the preliminary classification label of the geographic object Geo-Object as an entity, determining the node type of the geographic object knowledge graph; and expanding the nine intersection model based on dimensions performing spatial operation to extract initial topological relationship; extracting initial functional relationship based on weak supervision semantic initial rule; taking all calculated geometric features and spectral features as attributes of the geographic object Geo-Object entity, and finally outputting an initial geographic object knowledge graph basic triple set , wherein represents an attribute triple, represents a set operation, represents a head entity, represents a tail entity, represents a relationship. 4.The method of claim 3, wherein, The specific process of step S3 is: Step S31, uncertainty analysis and identification: for each triple in the middle triple set, the confidence of the triple is evaluated, and triples with confidence lower than a preset threshold or lacking semantic evidence are filtered out as triples to be enhanced. Step S31, uncertainty analysis and identification: for each triple in the middle triple set, the confidence of the triple is evaluated, and triples with confidence lower than a preset threshold or lacking semantic evidence are filtered out as triples to be enhanced. Step S32, set up GraphRAG depth retrieval mechanism: based on the head entity, tail entity and relationship type of the to-be-enhanced triple, construct a refined retrieval query, use the pre-constructed semantic graph of the external corpus, identify semantic clusters through Leiden algorithm, recall the entire semantic cluster related to the query target topic; perform multi-hop path analysis, mine cross-paragraph and cross-document implicit causality and semantic chain, sort the recalled text blocks by relevance, select the text context set with the most information quantity, and form an evidence chain ; Step S33, LLM-based structured knowledge extraction and completion: LLM analyzes the evidence chain , verifies the semantic rationality of the relationship in , performs entity extraction, relationship extraction, attribute extraction, and event extraction tasks, outputs formalized knowledge triples and event quadruples, identifies synonymous entities, unifies the identification, generates new functional relationship triples consistent with the common sense of the vertical field, and finally outputs the enhanced Geo-KG set . 5.The method of claim 4, wherein, In step S4, the tensor decomposition model is used. Training the Geo-KG knowledge graph yields the knowledge graph embedding vectors. The process is as follows: First, the enhanced Geo-KG set is used as the training set , where is the head entity, , is the entity set, is the relation, , is the semantic relation set, is the tail entity, ; Model selection and initialization is then performed, using tensor decomposition model As a knowledge graph embedding algorithm, initialize all entities and relation complex vector embeddings; by minimizing a scoring function Training is performed, the scoring function being defined as: ； wherein is the conjugate of is the real part extraction operator, and〈〉is the complex three-linear point product. After training, the knowledge graph embedding vectors of all categories of Geo-Object are obtained. , , For discrete entity indexing, For the first The latent feature vectors of each entity represent global semantic association information. For entity collection The base number. 6.The method of claim 5, wherein, In the step S4, the probabilistic topology intensity matrix is obtained based on the large-scale remote sensing true value data set statistics and quantization. The process is as follows: First, use the ground truth label dataset of remote sensing images. Define the set of land cover categories as , , For the first Each category of land cover It represents the total number of land cover categories; Then, spatial co-occurrence frequency statistics are performed, which includes pixel pair statistics, total pixel pair statistics, and calculation of the initial co-occurrence frequency. Pixel pair statistics are... Each pair of land cover categories , For the first Each category of land cover For the first Each land cover category, statistics and The number of spatially adjacent or functionally symbiotic pixel pairs Total pixel pair statistics are statistics Number of all pixel pairs appearing ; Calculate the first co-occurrence frequency in the initial co-occurrence frequency matrix Line number Column elements : ； Then, the probabilistic topology intensity value is calculated, and the LLM and Geo-Object geometry factors are introduced to correct the final intensity value The three factors are fused by a Sigmoid function The normalized value is obtained: ； wherein, is the weighted fusion score, defined as: ； in, For geometric feature factors based on Geo-Object, Based on Statistical factors of the knowledge graph structure. This serves as a verification factor for the semantic rationality of topological relationships based on LLM. All are weighted coefficients; Finally, the probabilistic matrix is outputted, obtaining a probabilistic topology strength matrix with dimensions of , as a soft constraint of topology consistency for the downstream model. 7.The method of claim 6, wherein, In the step S5, the conflict knowledge is searched by The process of searching the evidence chain for the conflict knowledge is as follows: First, the following are inputted: and the non-revocable primary knowledge source, which is a hard rule base.​ Then, semantic embedding verification, topological intensity verification and domain logic verification are respectively performed to identify conflicts; Finally, all identified conflicts are marked as a set of triples to be corrected. ,use right For each conflicting triple, a chain of evidence is retrieved. The goal is to obtain a structured, traceable chain of evidence. Chain of evidence Points to the most authoritative text blocks or logical reasoning chains in an external corpus, used to support or refute conflicting knowledge. 8.The method of claim 7, wherein, In step S5, the goal of semantic embedding verification is to ensure that the semantic relationships between geo-objects conform to domain common sense. The process of semantic embedding verification is as follows: evaluation and Vector distance between ,like Less than the preset threshold However, the definition of a first-level knowledge source and If they must be independent, then they are marked as semantic conflicts; The goal of topology strength check is to ensure that spatial coexistence flexibility complies with hard rules. The process of topology strength check is to check the strength value of . If the primary knowledge source defines and never coexist, but the check program finds greater than the set threshold , it is marked as a topology conflict. The goal of domain logic verification is to ensure that the graph structure conforms to the domain axioms. The process of domain logic verification is to perform rule-based logical inference on the triples in , and if there exists in , while the first-level knowledge source defines , which is impossible to be contained in , then it is marked as a logic conflict. 9.The method of claim 8, wherein, In step S5, the process of LLM combining knowledge summarization and conflict resolution strategies to adjudicate conflicts is: First, the LLM receives , primary knowledge source rules, and evidence chains ; Then, LLM executes the following knowledge summarization and conflict resolution strategies: Credibility assessment: based on the authority of the evidence source Credibility score; Repeat entity disambiguation: checking again whether the entities in the conflict triples have the problem of synonymy and non-merging; Relationship consistency verification: checking whether the relationship logic of the conflict triples conforms to the domain axioms and authoritative evidence; Finally, LLM correction Incorrect semantic embedding values ​​or topological strength values ​​are used to generate the final map set. . 10.The method of claim 9, wherein, In the step S5, finally, write into the Neo4j graph database, output the vertical field remote sensing feature knowledge graph corresponding to the field or project, including , , , as external knowledge, reconfigure semantic clusters, improve the recall accuracy of the next round , and form a closed-loop knowledge mining system.

Images